PaperlayerView exact revision on arXiv

Source-linked AI summary

FedDermaSeg: Federated Learning for Dermatological Image Segmentation

Anabik Pal, Ganesh Patidar, Bikash Santra

arXiv:2610.08574v1cs.CVcs.AIcs.LG

TL;DR

Federated skin-lesion segmentation offers a way to train collaboratively without centralizing medical images, addressing privacy concerns in centralized training. The paper evaluates a FedAvg model on simulated ISIC 2018 clients and reports performance comparable to centralized training and better than local models.

  • Problem

    Centralized segmentation training aggregates medical images and masks, raising privacy concerns, while evidence for federated skin-lesion segmentation remains limited.

    Source: §1 · p. 2

  • Method

    The paper trains a FedAvg skin-lesion segmentation model using a simulated five-client split of ISIC 2018 training data, then evaluates it on ISIC 2018 test and PH2 data.

    Source: §1 · p. 2, §3.2 · p. 5

  • Results

    The federated global model performs better than the best local model and comparably to centralized training, with TransUNet more suitable than UNet for skin-lesion segmentation.

    Source: §4 · pp. 7–8, §5 · p. 10

  • Takeaways & Limitations

    Federated learning can improve segmentation performance in data-restricted scenarios without requiring centralized access to client images.

    Source: §1 · p. 2, §5 · p. 10

    The evaluation uses randomly split simulated clients, and future work targets more nuanced non-IID conditions.

    Source: §3.2 · p. 5, §5 · p. 10

Abstract

from arXiv · show

Skin cancer is a major global health concern, and early detection and accurate lesion delineation are important for effective diagnosis and treatment planning. Automated skin lesion analysis can assist dermatologists, with lesion segmentation serving as a fundamental step in computer-aided diagnostic systems. Conventional deep learning-based segmentation models typically rely on centralized training, where images and their corresponding segmentation masks are collected on a central server. Such data aggregation raises privacy concerns in medical applications and requires substantial centralized computational resources. To address these limitations, we investigate the feasibility of federated learning for privacy-preserving skin lesion segmentation. The training and validation sets of the ISIC 2018 Skin Lesion Segmentation Challenge dataset are used to simulate a distributed learning environment and develop a federated segmentation model. The resulting model is evaluated on the ISIC 2018 test set and the PH2 dataset to assess its performance and generalizability. Experimental results demonstrate that the federated model achieves performance comparable to centralized training while consistently improving upon the locally trained models. These findings demonstrate the potential of federated learning for collaborative skin lesion segmentation without requiring centralized aggregation of medical images.

1 Introduction

The paper motivates federated skin-lesion segmentation as a way to collaborate without centralizing sensitive images, addressing limited evidence for federated segmentation. It develops and evaluates a FedAvg-based model and examines segmentation networks and loss functions.

1 Introduction

  • Skin-lesion segmentation is presented as a prerequisite for automated systems that could assist early detection and reduce experts’ workload.The introduction notes limits in classical image-processing methods and motivates deep-learning approaches.Source: §1 · p. 1
  • Federated learning trains a shared model across collaborating sources without transferring raw images, addressing privacy concerns and decentralizing computation.A parameter server broadcasts weights, aggregates client updates with weighted averaging, and repeats the process until convergence.Source: §1 · p. 2
  • The paper develops a cross-silo FedAvg skin-lesion segmentation model and evaluates it on ISIC 2018 test data and PH2 after training on ISIC 2018.The study simulates clients and a parameter server, gives clients equal aggregation weights, and examines performance trade-offs.Source: §1 · p. 2
  • The study also compares commonly used deep segmentation networks and segmentation loss functions.The introduction frames this work against limited prior attempts to develop federated skin-lesion segmentation models.Source: §1 · pp. 2–3

2 Methodology

The method trains segmentation models locally at clients, aggregates their updates into a global model with equal client weights, and repeats this process over communication rounds. It evaluates UNet and TransUNet with several loss functions.

Federated training

  • Each client trains a received model on local data for E updates, then sends its weights to the server for equal-weight FedAvg aggregation over T communication rounds.The aggregated global model is broadcast back to clients for the next round.Source: §2 · pp. 3–4

Segmentation architectures

  • UNet uses convolutional and deconvolutional layers with local interactions, while TransUNet adds transformer layers for global interactions between image features.Both architectures are trained with the federated algorithm and compared.Source: §2 · p. 4

Loss functions

  • The study tests binary cross-entropy, Dice, and Focal losses because lesion regions are smaller than normal-skin regions and BCE can favor normal regions.The authors state that Dice and Focal losses have some capability to address this bias.Source: §2 · p. 4
  • Dice and Focal loss definitions use ground-truth and predicted masks, with ε specified to address division by zero.The paper defines y as the ground-truth mask, ŷ as the predicted mask, and ε as a small number.Source: §2 · p. 5

3 Experimental Setup

The experimental setup uses public ISIC 2018 and PH2 dermoscopic datasets, simulates five clients by splitting ISIC training data, and evaluates segmentation with common metrics.

3.1 Datasets

  • ISIC 2018 provides 2,594 training, 100 validation, and 1,000 test images, while PH2 provides 200 annotated images across three lesion categories.The ISIC images span diverse clinical conditions, and PH2 includes nevus, atypical, and melanoma categories.Source: §3.1 · p. 5

3.2 Data Simulation for FL Setup

  • The ISIC 2018 training set is split into five disjoint client datasets of 518 or 519 images, with random splitting hypothesized to introduce some non-IID distribution.Validation assesses local and federated updates, while ISIC test and PH2 data evaluate the trained model.Source: §3.2 · pp. 5–6

3.3 Implementation Details and Performance Metrics

  • UNet and TransUNet are implemented with SGD at learning rate 0.001 and momentum 0.9, and performance is evaluated using Dice, IoU, accuracy, precision, recall, and F1.The number of local interactions and total rounds are selected experimentally; the setup compares local and centralized models as well.Source: §3.3 · p. 6

4 Results and Discussions

The experiments compare federated, local, and centralized segmentation, then test generalization and training choices. Across the reported datasets, global-FL is comparable to global-L and outperforms the best local model.

  • For both UNet and TransUNet, global-FL outperforms the best local model and performs comparably to global-L.Source: §4 · pp. 7–8
  • For most ISIC 2018 images, global-FL and global-L are difficult to distinguish visually, though the third selected image is poorly segmented.Source: §4 · p. 8, Figure 2 · p. 8
  • On PH2, performance degrades for all ISIC-trained models, but global-FL remains much better than the best local model and comparable to global-L.Source: §4 · p. 9
  • Dice + Focal performs superiorly or similarly to other tested losses and is selected for implementation.Source: Figure 3 · p. 8, §4 · pp. 9–10
  • With 1, 5, or 10 epochs per federated iteration, one epoch performs best on ISIC 2018, despite higher computational expense.Source: Figure 4 · p. 9, §4 · p. 10

5 Conclusion and future scope of Work

The conclusion reports that federated training improves segmentation in data-restricted scenarios, while usually trailing centralized training. It identifies noisy images and several directions for future work.

5 Conclusion and future scope of Work

  • The federated global model improves segmentation in data-restricted scenarios but usually trails centralized training; TransUNet outperforms UNet.Source: §5 · p. 10
  • Hair artifacts make dermatological images noisy and can mislead training, so reliable models for such images remain immediate future work.Source: §5 · p. 10
  • Future work includes better networks and visual foundation models, lesion-attribute segmentation, nuanced non-IID partitions, and privacy-preserving aggregation security.Source: §5 · p. 10
Loading 2610.08574v1…